Interpretability-Mask: a label-preserving data augmentation scheme for better classification

Data augmentation effectively alleviates the over-fitting problem in convolutional neural network-based (CNN-based) models, especially in the limited dataset. However, the inconsistency problem between the augmented sample and its original label is still a critical challenge during the augmentation operation. In this paper, we propose a novel data augmentation scheme named Interpretability-Mask (IM), which exploits the interpretability of the classifier to obtain the most discriminative regions and preserve label invariance. Concretely, we first construct a set-based representation for a sample and its label by superpixel segmentation and the local interpretable model-agnostic explanations (LIME) operator. Secondly, the sample represented by the superpixel set is utilized to synthesize the region-level disturbance augmentation sample with a random removal strategy. Then, the label constructed by the most interpretive superpixel set is applied to maintain the consistency between the augmented sample and its original label. Lastly, the augmentation scheme will be randomly used to the training stage. Extensive experiments are conducted on challenging datasets. A significant improvement in classification performance has achieved with the IM scheme. On the CIFAR-10 dataset, the Top-1 error rate drops by 2.15% at most. On the CIFAR-100 dataset, the Top-1 error rate decreases by up to 3.69%. And the maximum decline of the Top-1 error rate is 3.35% on the Mini-ImageNet. Experimental results manifest the effectiveness and generality of the proposed method.

Automated summary

Data augmentation effectively reduces overfitting in CNN-based models, especially with limited datasets. This paper proposes a novel data augmentation scheme called Interpretability-Mask (IM), which leverages the interpretability of the classifier to identify discriminative regions and preserve label consistency. By constructing a set-based representation using superpixel segmentation and the LIME operator, the IM scheme synthesizes region-level augmented samples while maintaining consistency with the original labels. Extensive experiments demonstrate the effectiveness and generality of the proposed method, achieving significant improvements in classification performance.